Time series analysis of image metadata The data set that I have—a user's corpus of photos with time taken and an array of locations in each of them. Location is a bit hierarchical in nature, thus an array, like country, state, county, place.
Between two users, I'm trying to find an image from user A which is closest(most relevant recommendation on the basis of time and location) to one from user B. 
The approach I'm taking is to convert each photo of user A(and B) into a vector with one coefficient representative of time and one representative of location and then choose between Euclidean distance or hamming distance to find the closest.
I need a few suggestions here:


*

*I'm currently deriving the time coefficient using the following method. Plot images taken on each day against time. The image coefficient of each images is the number of images taken on the day/daily average
I need alternative recommendations, like standard deviation, or any thing else. The intuition is, that image relevant in terms of time (which it is if the user has taken higher than average number of images on a day).

*For location, I'm doing a similar thing, calculating the importance of a location, by number of times the particular location is observed/total number of locations. This I'm doing for all the elements in the array. Now how do I combine these numbers to get one representative number? Is there a different formula I should be using?

*Finally, assuming that we have a vector of time and location, is euclidean distance a better representation than hamming distance?
Any help with the thinking will be highly appreciated.
 A: Let me assume from your question "Between two users, I'm trying to find an image from user A which is closest(most relevant recommendation on the basis of time and location) to one from user B." that the content of the photo is irrelevant from the scoring.
You could use simple least square regression using dummy variables for each dates and location, sort the coefficients and classify individual by their highest few coefficients.
Most locations and dates that are not included in the data set will be 0, and should not appear in the model, and at most you will have n parameters for n locations (this can be also grouped as street - town - country).
For the 'importance of location' by frequency of the photo, the dummy variable coefficient will be adjusted accordingly when you include an interaction term, such as (time * location). Only take coefficients, and
Alternatively, you could implement a classification method according to scoring? 
This way will be less exact in matching coinciding events, such as 2016 Christmas eve in Vegas. However, this way you can match similarities between the 'days of week' 'time of the day' such as Friday evenings can be isolated.
First method would be more appropriate for suggesting a restaurant that could be available to both parties, finding a person who was in an event at certain day as we have more resolution.
The second method can be used in context where the matching is for somewhat  broader preference or personal patterns for example date matching.
Hope my 2 cents helps!!
